Risks · Established · Intermediate

Model Drift

The gradual decay of model performance as the world changes away from the data it was trained on.

What Model Drift is

Data drift is a change in input distribution; concept drift is a change in the relationship between inputs and outcomes. Both degrade accuracy without any code changing.

How it works

Monitoring compares live feature distributions and prediction mixes against training baselines, and tracks realised outcomes where labels eventually arrive. Retraining is triggered on thresholds or a schedule.

Why it matters

It is the reason 'we deployed it and it worked' is never the end of the story, and the most common cause of silent value loss.

Common uses

→Production monitoring
→Retraining triggers
→Model risk management

Watch for

✓Label delay hides drift
✓Alert thresholds are hard to tune

Continue exploring

MLOps (2)AI Observability Model Evaluation Feature Store